It contains 16,288 pass/fail tests across 2,937 pages from 2,343 documents in 92 languages. Each test is traced to its source and can be queried by content tags such as tables, equations, handwriting, rotation, small text, scripts, and languages.
The benchmark reports a headline score as the mean of text, tables, and layout family scores. Datalab says its scorer is open code and does not use model judging. Its published table lists Gemini 3.8 Flash at 92.9, Datalab accurate at 92.7, Reducto at 90.8, LlamaParse at 90.4, Mistral OCR at 89.5, and Tesseract at 17.6. The release also includes latency and price-per-1,000-pages comparisons.
The dataset and code are open, and users can recompute scores for selected tags or their own runs. The source discloses that Gemini 3.8 Flash and Claude Sonnet 5.5 were used in creating and verifying at least some tests. The publication does not establish an independent ranking of parsers, nor does it show performance on any particular organization's document corpus.